Parallelism - Cornell UniversityParallelism across iterations •Try to compute multiple iterations...

ParallelismCS6787 Lecture 8 — Fall 2017

So far

• We’ve been talking about algorithms

• We’ve been talking about ways to optimize their parameters

• But we haven’t talked about the underlying hardware• How does the properties of the hardware affect our performance?• How should we implement our algorithms to best utilize our resources?

What does modern ML hardware look like?

• Lots of different types• CPUs•GPUs• FPGAs• Specialized accelerators

• Common thread: all of these architectures are highly parallel

Parallelism: A History

• The good old days: if I want my program to run faster, I can just wait• Moore’s law — number of transistors on a chip doubles every 18 months• Dennard scaling — as transistors get smaller, power density stays constant

• This “free lunch” drove a wave of innovation in computing• Because new applications with bigger data were constantly becoming feasible• Drove a couple of AI boom-bust cycles

• But also drove a lack of concern for systems efficiency• Why work on making efficient systems when I can just wait instead?

Moore’s Law: A Graphic

The End of the Free Lunch

• In 2005, Herb Sutter declares — “The Free Lunch Is Over” and that there will be “A Fundamental Turn Toward Concurrency in Software”• He’s not the only one that was saying this.

• You can see this on the previous figure as trends start to flatten out.

• Why? Power• Dennard scaling started breaking down — no longer fixed power/area• Too much heat to dissipate at high clock frequencies — chip will melt

The Solution: Parallelism

• I can re-write my program in parallel

• Moore’s law is still in effect• Transistor density still increasing exponentially

• Use the transistors to add more parallel units to the chip• Increases throughput, but not speed

The Effect of Parallelism

• Pros:• Can continue to get speedups from added transistors• Can even get speedups beyond a single chip or a single machine

• Cons:• Can’t just sit and wait for things to get faster• Need to work to get performance improvements• Need to develop new frameworks and methods to parallelize automatically

What benefits can we expect

• If we run in parallel on N copies of our compute unit, naively we would expect our program to run N times faster

• Does this always happen in practice?

• No! Why?

Amdahl’s Law

• Gives the theoretical speedup of a program when it’s parallelized

• Slatency is total speedup• p is the parallelizable portion of the algorithm• s is the number of parallel workers/amount of parallelism

From Wikipedia

Amdahl’s Law (continued)

From Wikipedia

Consequences of Amdahl’s Law

• Diminishing marginal returns as we increase the parallelism

• Can never actually achieve a linear or super-linear speedup as the amount of parallel workers increases

• Is this always true in practice?

• No! Sometimes we do get super-linear speedup. When?

What does modern hardware look like?

• CPUs• Many parallel cores• Deep parallel cache hierarchies — taking up most of the area• Often many parallel CPU sockets in a machine

• GPUs• Can run way more numerical computations in parallel than a CPU• Loads of lightweight cores running together

• In general: can run many heterogeneous machines in parallel in a cluster

Sources of parallelismFrom most fine-grained to most course-grained

On CPUs: Instruction-Level Parallelism

• How many instructions in the instruction stream can be executed simultaneously?

• For example:• C = A * B• Z = X * Y• S = C + Z

• Important for pipelining, and used fully in superscalar processors.

The first two instructions here can be executed in parallel

On CPUs: SIMD/Vector Parallelism

• Single-Instruction Multiple-Data• Perform the same operation on multiple data points in parallel

• Uses registers that store and process vectors of multiple data points• Latest standards use 512-bit registers, which can hold 16 floating point numbers

• A long series of instruction set extensions for this on CPUs• SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2, AVX, AVX2, AVX-512, …

• Critical for dense linear algebra operations common in ML

On CPUs: Multicore Parallelism

• Modern CPUs come with multiple identical cores on the same die

• Cores can work independently on independent parallel tasks• Unlike ILP and SIMD

• Cores communicate through shared memory abstraction• They can read and write the same memory space• This is done through a sophisticated cache hierarchy

• Significant cost to synchronize multiple CPUs working together

On CPUs: Multi-socket parallelism

• Modern motherboards have multiple sockets for CPUs

• Cores on these CPUs still communicate through shared memory

• But latency/throughput to access memory that is “closer” to another CPU chip is worse than accessing your own memory

• This is called non-uniform memory access (NUMA)

On GPUs: Stream Processing

• Given a stream of data, apply a series of operations to the data• Operations are called kernel functions

• This type of compute pattern is well-suited to GPU computation• Because compared with CPUs, GPUs have much more of their area devoted

to arithmetic but much less devoted to memory and caches

• There’s additional parallel structure within a GPU• For example, in CUDA threads running the same program are organized into

warps and run at the same time

On specialized accelerators and ASICs

• Whatever you want!

• The parallelism opportunities are limited only by the available transistors

• We will see many new accelerators for ML with different parallel structures and resources• Some will look like FPGAs: e.g. CGRAs• Some will just speed up one particular operation, such as matrix-matrix multiply

The Distributed Setting

• Many workers communicate over a network• Possibly heterogeneous workers including CPUs, GPUs, and ASICs

• Usually no shared memory abstraction• Workers communicate explicitly through passing messages

• Latency much higher than all other types of parallelism• Often need fundamentally different algorithms to handle this

How to use parallelism in machine learningFrom most fine-grained to most course-grained

Recall

• Stochastic gradient descent

• Can write this as an algorithm:

• For t = 1 to T• Choose a training example at random• Compute the gradient and update the model• Repeat.

xt+1 = xt � ↵rf(xt; yit)

How to run SGD in parallel?

• There are several places where we can extract parallelism from SGD.

• We can use any or all of these places• Often we use different ones to correspond to the different sources of parallelism

we have in the hardware we are using.

Parallelism within the Gradient Computation

• Try to compute the gradient samples themselves in parallel

• Problems:• There usually is not much work here, so not much work to split up• We run this so many times, we will need to synchronize a lot

• Typical place to use: instruction level parallelism, SIMD parallelism

xt+1 = xt � ↵rf(xt; yit)

Parallelism with Minibatching

• Try to parallelize across the minibatch sum

• Problems:• Still run this so many times, we will need to synchronize a lot• Can have a tradeoff with statistical efficiency, since too much minibatching

can harm convergence

• Typical place to use: all types of parallelism

xt+1 = xt �↵

B

BX

b=1

rf

�xt; yib

�

Parallelism across iterations

• Try to compute multiple iterations of SGD in parallel• Parallelize the outer loop — usually a good idea

• Problems:• Naively, the outer loop is sequential, so we can’t do this without fine-grained

locking and frequent synchronization

• Typical place to use: multi-core/multi-socket parallelism

xt+1 = xt � ↵rf(xt; yit)xt+1 = xt � ↵rf(xt; yit)xt+1 = xt � ↵rf(xt; yit)xt+1 = xt � ↵rf(xt; yit)

Parallelism for hyperparameter optimization

• Just run multiple copies of the whole algorithm independently, and use them to do hyperparameter optimization

• Problems:• Can’t do this if you don’t want to do hyperparameter optimization• Isn’t actually useful once you’ve already set your parameters

• Typical place to use: distributed computation

What about our other methods?

• We can speed up all our methods with parallel computing• Minibatching — has a particularly close connection with parallelism• SVRG• Momentum

• And any SGD-like algorithm lets us use the same ways to extract parallelism from it• Things like gradient descent, stochastic coordinate descent, stochastic gradient

Langevin dynamics, and many others.

Asynchronous Parallelism

Limits on parallel performance

• Synchronization• Have to synchronize to keep the workers aware of each other’s updates to the

model — otherwise can introduce errors

• Synchronization can be very expensive• Have to stop all the workers and wait for the slowest one• Have to wait for several round-trip times through a high-latency channel

• Is there something we can do about this?

Idea: Just Don’t Synchronize

• Not synchronizing adds errors due to race conditions

• But our methods were already noisy — maybe these errors are fine

• If we don’t synchronize, get almost perfect parallel speedup

Fast Parallel SGD: HOGWILD!Multiple parallel workers

Asynchronous parallel updates (no locks) to a single shared model

xt

Questions?

• Upcoming things• Paper Review #7 — due today• Paper Presentation #8 on Wednesday — HOGWILD!